RATIONAL SERIES IN THE FREE GROUP AND THE CONNES OPERATOR
AARON LAUVE AND CHRISTOPHE REUTENAUER
Abstract. We characterize rational series over the free group by using an oper- ator introduced by A. Connes. We prove that rational Malcev–Neumann series posses rational expressions without simplifications. Finally, we develop an effec- tive algorithm for solving the word problem in the free skew field.
1. Introduction Fix a field k and a finite set of indeterminants X. The free commutative field k(X) is defined as the initial object in the category of (commutative) fields F with embeddings k[X]֒→F . It also has a well understood realization as the field of rational functions: k(X)= f/g | f, g ∈ k[X], g = 0 . The focus of this paper is the noncommutative counterpart to k(X), the free (skew) field k<(X>.) It has an analogous categorical definition involving embeddings of k X , but one can hardly say that k<(X>) is well understood. As an illustration, the reader may take a moment to verify that x − z−1 1 − yx −1 y − z + y−1 − z−1 1 − x−1y−1 −1 x−1 − z = 0 (1.1) without allowing the variables to commute. A commonly used realization of the free field, due to Lewin [Lew74], involves Malcev–Neumann series over the free group Γ(X) generated by X. These are noncommutative, multivariate analogs of Laurent series; we recall the details in Section 2. While the Lewin realization of the free field is powerful, it suffers from an incon- venient asymmetry. We illustrate with an identity of Euler [Car00]: xk = 0. Z k∈ −k k This looks absurd, but rewriting it as k≥0 x + x k≥0 x , we recognize the geometric series expansions of 1/(1 − x−1) and x/(1 − x), respectively. Adding these fractions together does indeed give zero. Of course, kJX ∪ X−1K is not a ring, so the mixing of series in x and x−1 is usually disallowed. Nevertheless, passing through steps such as this one is a useful technique for proving identities in the free field. (See Appendix A.)
Date: July 4, 2012. 1 2 AARONLAUVEANDCHRISTOPHEREUTENAUER
In this paper, we give Euler’s identity firm footing in the noncommutative setting. (See [BHS09, BR03] for more on the commutative setting.) Our main result in this vein is a characterization of which series over the free group Γ(X) have rational expressions (Theorem 3.1 of Section 3). The main ingredient in the proof is a Fredholm operator F on the free group that we describe in Section 2.3. This operator featured prominently in a conjecture of Alain Connes [Con94] that was proven in [DR97]. Our result extends the main result there. Briefly, an element a ∈ kΓ is rational if and only if its associated Connes operator [F, a] has finite rank. In this paper, we also tackle the problem of rendering a given rational expression into Malcev–Neumann form (Section 4). Our main result here, Theorem 4.9, is an effective algorithm for solving the word problem in k<(X>) . Recall that the word problem for free groups (when are two words equal in Γ(X)?) is decidable, even though this is famously not the case for all groups [Nov55, Boo58]. The analogous problem for the free field (when are two expressions equal in k<(X>) ?) was first considered in [Coh73] and taken up again in [CR99]. Here is a simple example: Are x−1 (1 − x)−1 and x−1 + (1 − x)−1 equal? The answer is, “yes,” as the reader may easily verify. While algorithms are presented in [Coh73, CR99], their complexity seems very high. We derive our algorithm from Fliess’ proof [Fli71] of the following fact: simplifications in rational expressions preserve rationality. Here, at last, we may say that the word problem is certainly not undecidable. Though the complexity of our algorithm could certainly be improved.
2. The Free Field, Rational Series, and the Connes Operator 2.1. The free field. The free field k<(X>) is defined as the initial object in the category of epic k X skew fields with specializations. (See [Coh95, Ch. 4], which also contains a realization in terms of full matrices over k X .) In what follows, we use Lewin’s realization in terms of formal series. 2.1.1. Operations on series over the free group. Let Γ = Γ(X) denote the free group generated by a finite set of noncommuting indeterminants X. Let kΓ denote the vector space with basis Γ and kΓ denote the formal series over Γ, i.e., functions a: Γ → k, which we may write as ω∈Γ(a, ω) ω or ω∈Γ aω ω as convenient. The support supp(a) of a series a is the set of its nonzero coefficients, {ω ∈ Γ | (a, ω) = 0}. The sum a + b of two series over Γ is well-defined and defined by (a + b, ω) = (a, ω) + (b, ω) for all ω ∈ Γ. The Cauchy product a b (or ab) of two series is well-defined if (ab, ω) := (a, α)(b, β) α,β∈Γ αβ =ω is a finite sum for all ω ∈ Γ. We will also have occasion to use the Hadamard product a ⊙ b, defined by (a ⊙ b, ω) := (a, ω)(b, ω) for all ω ∈ Γ. We define now a unary operation called the star operation: given a ∈ kΓ, we let a∗ denote the sum 1+ a + a2 + , if this is a well-defined element of kΓ; in other words, if for any ω ∈ Γ, there are only finitely many tuples (ω1,...,ωn) such that RATIONALSERIESANDTHECONNESOPERATOR 3
Γ ω = ω1 ωn and (a, ω1) (a, ωn) = 0. In this case, it is the inverse of 1 − a in k under the Cauchy product. 2.1.2. Rational series over the free group. Fix a total order < on Γ compatible with its group structure. (See Appendix C for an elementary example.) The Malcev– Neumann series (with respect to <) is the subset k((Γ)) ⊆ kΓ of series with well- ordered support. That is, a ∈ k((Γ)) if every nonempty set A ⊆ supp(a) has a minimum element. One shows that k((Γ)) is closed under addition, multiplication and taking inverses, i.e., k((Γ)) is a skew field. See [Coh95, Ch. 2.4], [Pas85, Ch. 13.2], or [Sak09, Ch. IV.4]. The inverse of a Malcev–Neumann series a is built in a
geometric series–type manner: if a = aoωo + ωo<ω aωω, then −1 ∗ −1 −1 −1 −1 −1 a = 1 − aω ωωo aoωo = ao ωo aω ωωo , (2.1) ωo<ω ωo<ω −1 −1 −1 with aω = −aωao and ωo being the minimum element of supp (a ). Lewin [Lew74] showed that the free field k<(X>) is isomorphic to the rational closure of k X in k((Γ)), i.e., the smallest subring of k((Γ)) containing k X and closed under addition, multiplication and taking inverses. An alternative proof, using Cohn’s realization of the free field in terms of full matrices, appears in [Reu99]. 2.2. Rational series over free monoids. The rational series in k((Γ)) introduced above are a particular case of rational series over an alphabet A. In this more general notion, one starts with polynomials on the free monoid kA∗ and builds expressions with (+, ×, (·)∗). (Here, for an expression p ∈ kA∗ without constant term, p∗ := 1+ p + p2 + .) Indeed, let X = {x | x ∈ X} represent formal inverses for the elements of X, and put A = X ∪ X. Then (2.1) allows us to replace the (·)−1 operation over Γ with the (·)∗ operation over A. We do so freely in what follows and exploit results from the latter theory that we collect here. ∗ 2.2.1. Hankel rank. Let S = w(S, w)w be a series over A . Given a nonempty word u ∈ A∗, we define the new series u ◦ S ∈ k A , a (right) translate of S, by ◦ u S := (S, w)wo. w : w=wou Letting ε denote the empty word, put ε ◦ S := (S, ε) and extend bi-linearly to view ◦ as a map ◦ : kA∗ ⊗ k A → k A . (See [Sak09, Ch. III.4] or [BR11, Ch. 1.5], where the image is denoted Su−1.) The Hankel rank of the series S, introduced in [Fli74], is the rank of the operator ◦ S : kA∗ → k A . We extend this construction to series over the free group Γ in Section 3.4. The following is classical. Theorem 2.1 (Fliess). A series S ∈ k A over a free monoid A∗ is rational if and only if its right translates {f ◦ S | f ∈ kA∗} span a finite dimensional subspace of k A . An easy corollary is that the Hadamard product preserves rationality. For a proof, see [BR11, Th. 1.5.5]. 4 AARONLAUVEANDCHRISTOPHEREUTENAUER
Theorem 2.2 (Sch¨utzenberger). If S, T are rational series, then S ⊙ T is rational. 2.2.2. Recognizable series. We recall some additional results on recognizable series that will be useful in Section 4. Fix a series S = w(S, w)w ∈ k A . Let L = supp(S). The language L is said to be recognizable if there is an au- tomaton1 that accepts L and no other words. For example, the machine in Figure 1 accepts supp(b + ac)∗; the input state is marked with an arrow and the output state is doubly circled. More generally, we may extend the notion of automata so that a b 0 1 c
Figure 1. An automaton recognizing the language supp(b + ac)∗. edge-labels carry coefficients other than 0 or 1, e.g., taking values in some semiring k, and speak about S being recognizable by a k-automaton. Theorem 2.3 (Sch¨utzenberger). A series S is rational iff it is recognizable. We say that S is representable, with order n, if there exist row and column vectors λ ∈ k1×n, ρ ∈ kn×1, and a monoid morphism : A∗ → kn×n satisfying (S, w)= λ (w)ρ for all w ∈ A∗. For example, a representation (λ, , ρ) of the series (b + ac)∗ is provided by 0 1 1 0 0 0 1 λ = 1 0 , (a)= , (b)= , (c)= , ρ = . 0 0 0 0 1 0 0 x Here, (x)ij = 1 iff there is an edge i −→ j in the automaton. The equivalence between recognizable and representable series is well-known. (In fact, the usage of “recognizable” in the literature for what we call here “representable” is common- place.) In [Fli74], Fliess shows that the least n for which S has a representation is the Hankel rank of S [BR11, Th. 2.1.6]. 2.3. The Connes operator. In his book Noncommutative Geometry, Connes [Con94] gives a new proof of a celebrated C∗-algebra result2 of Pimsner–Voiculescu [PV82] using the machinery of “Fredholm modules.” We recount the details that are relevant to the present work. 2.3.1. Reduced words. Recall that each element ω ∈ Γ has a unique reduced expres- −1 sion ω = x1 xn with xi ∈ X or xi ∈ X for all i and no pair (i, i+1) satisfying xixi+1 = 1Γ. We say that ω has length ℓ(ω)= n. ∗ We identify Γ with the subset of reduced words, red(X ∪ X) , within the free ∗ monoid: say ω ∈ (X ∪ X) is reduced if it contains no factor of the form xx or x x
1A finite state machine that processes words and transitions between states by reading letters one at a time [Sak09]. ∗ 2Namely, the reduced C -algebra of the free group does not contain nontrivial idempotents. RATIONALSERIESANDTHECONNESOPERATOR 5 for any x ∈ X. We often denote the inverse of an element ω ∈ Γ by ω, and put ω = ω, to reduce clutter in the expressions that follow. Finally, if ω ,...,ω are elements of . 1 r Γ, we write ω = ω1 ωn if the product is a reduced factorization of ω (in particular, each ωi is reduced). That is, ω = ω1 ωr in Γ and ℓ(ω)= ℓ(ω1)+ + ℓ(ωr). −1 Let ω = x1 xn (xi ∈ X or xi ∈ X) be the reduced expression for ω ∈ Γ\{1}. Its longest proper prefix ℘(ω) is the word x1 xn−1. Its (nonempty) suffixes comprise the set S(ω) = {xn, xn−1xn,...,x1 xn−1xn}. We take the codomains ∗ of ℘(·) and S(·)tobeΓor red(X ∪ X) , as convenient, with context dictating which is intended. 2.3.2. The Connes operator. Let G = (V, E) be the Cayley graph of Γ, the infinite tree with vertex set V = Γ and edge set E satisfying {υ,ω}∈E if and only if υ = ℘(ω) or ω = ℘(υ). (In this case, note that {αυ, αω} is again an element of E, for any α ∈ Γ.) On the linear space kG = kΓ ⊕ kE, define a bilinear operation Γ G Γ : k ⊗ kG → k as follows. For a = α∈Γ(a, α)α ∈ k and {υ,ω}∈E, put a ω := (a, α) αω and a {υ,ω} := (a, α) {αυ, αω}. α ∈Γ α ∈Γ Recall that a Fredholm operator between two normed vector spaces X and Y is a bounded linear map with finite dimensional kernel and cokernel. The Fredholm operator F: kG → kG that Pimsner–Voiculescu and Connes use is defined as follows. For all υ ∈V and {℘(ω),ω}∈E, 0 if υ = 1 F({℘(ω),ω}) := ω and F(υ) := {℘(υ), υ} otherwise. The reader may check that ker F = coker F = spank{1}. Definition 2.4. Given a ∈ kΓ, the Connes operator [F, a]: kG → kG is defined as the commutator F ◦ a − a ◦ F. We develop a formula for the action of [F, a] on an edge {℘(ω),ω}: [F, a] ℘(ω),ω = F ◦ a ◦ ℘(ω),ω − a ◦ F ◦ ℘(ω),ω = F aα α℘(ω), αω − a ω α∈Γ
= aα α℘(ω)+ aα αω − aααω + aααω α∈Γ α∈Γ α∈Γ α∈Γ ω∈ S(α) ω ∈ S(α) ω∈ S(α) ω∈ /S(α)
= aα α ℘(ω) − ω . (2.2) α∈Γ ω∈ S(α) Example 2.5. [F,xxy]{y,¯ y¯x¯} = xxy(¯y − y¯x¯)= xx − x, while [F,xyy]{y,¯ y¯x¯} = 0. After his proof of the Pimsner–Voiculescu result, Connes offers a conjecture about the rank of his operator [Con94, p. 342, Remark 3]. This was later proved by 6 AARONLAUVEANDCHRISTOPHEREUTENAUER
Duchamp and the second author [DR97]. A discrete topology analog of the conjec- ture goes as follows. Theorem 2.6. An element a of the Malcev–Neumann series k((Γ)) belongs to the rational closure of kΓ in k((Γ)) ⇐⇒ the operator [F, a] is of finite rank. That is, if the image [F, a]kG ⊆ kG is finite dimensional. This appears as Th´eor`eme 12 in [DR97], but the proof of the forward implication is not given. A complete proof is given here, as it follows from the proof of our first result below.
3. Rank of the Connes Operators By the rational closure of kΓ in kΓ, we mean the smallest subspace of kΓ that contains kΓ and is closed under the Cauchy product and the star operation. Here, we insist only that all products and stars involved are well-defined in the sense of Section 2.1.1. Our first result is the following generalization of Theorem 2.6. Theorem 3.1. A formal series a ∈ kΓ belongs to the rational closure of kΓ in kΓ ⇐⇒ the operator [F, a] is of finite rank. Sections 3.1 and 3.2 establish the necessary machinery for the proof. Sections 3.3 and 3.4 establish the forward and reverse implications, respectively. (See Appendix B for possible refinements and extensions.) As a first simplification, note that [F, ·]: kG → kG has finite rank if and only if F · Γ the restricted operator [ , ] kE : kE → k does. We focus on kE in what follows. 3.1. Elementary identities on Connes operators. Given a, b ∈ kΓ, it will be useful to express the operators [F, ab] and [F, a∗] in terms of [F, a] and [F, b].
Proposition 3.2. Suppose a, b ∈ kΓ are such that ab and a∗ are well-defined. Then [F, ab] and [F, a∗] are well-defined operators; moreover [F, ab] = [F, a]b + a[F, b], (3.1) [F, a∗] = a∗[F, a]a∗. (3.2) Proof of Identity (3.1). First, some seemingly inoccuous algebra: [F, ab]= Fab − abF = (Fab − aFb) + (aFb − abF) = [F, a]b + a[F, b]. Supposing a, b, ab ∈ kΓ, the operators [F, a], [F, b], [F, ab]: kE → kΓ are well-defined, in particular Fab and abF are well-defined operators by themselves. In order to conclude that the left- and right-hand sides of (3.1) are equal, we must check that aFb is also well-defined as an operator. This is handled in the next lemma. Lemma 3.3. If a, b, and ab belong to kΓ, then aFb is a well-defined operator from kE to kΓ. RATIONALSERIESANDTHECONNESOPERATOR 7
Γ Proof. Write a = α∈Γ aαα and b = β∈Γ bββ. Since ab ∈ k , we have ab = aα bβ γ, (3.3) γ∈Γ α,β∈Γ γ =αβ with γ=αβ aα bβ involving finitely many nonzero terms for each γ ∈ Γ. The effect of aFb on ℘(ω),ω ∈E is as follows: aFb ℘(ω),ω = aF bβ β℘(ω),βω β∈Γ
= a bβ β℘(ω)+ bβ βω β∈Γ β∈Γ ω∈ S(β) ω ∈ S(β)
= aαbβ αβ℘(ω)+ aαbβ αβω . (3.4) α,β∈Γ α,β∈Γ ω∈ S(β) ω ∈ S(β)
We claim that no υ ∈ Γ appears infinitely often in (3.4). Indeed, suppose {(αi, βi)} is an infinite sequence satisfying υ ∈ αi βi ℘(ω), αi βi ω for all i. Then there is an infinite subsequence {(αj, βj)} satisfying ∀j, υ = αj βj ℘(ω) or ∀j, υ = αj βj ω. The two cases are similar; we consider the latter. Let υ′ω denote the common value of ′ the elements of the subsequence, so αjβj = υ (∀j). Deduce from (3.3) that only a
finite number of the αj βj appear with nonzero coefficient aαj bβj . This proves the claim and the Lemma. Proof of Identity (3.2). If a ∈ kΓ is such that a∗ is well-defined, then a∗a, aa∗, a∗a∗ and a∗aa∗ are also well-defined, and (1 − a)a∗ = a∗(1 − a) = 1. Thus we may write [F, a∗] = Fa∗ − a∗F = a∗(1 − a)Fa∗ − a∗F(1 − a)a∗ = a∗Fa∗ − a∗aFa∗ − a∗Fa∗ + a∗Faa∗ = a∗[F, a]a∗. The above observations and Lemma 3.3 guarantee that the operators a∗Fa∗, a∗aFa∗ and a∗Faa∗ used in the intermediate steps are well-defined. Conclude that (3.2) holds whenever a∗ is well-defined. 3.2. Closed subspaces of kΓ. The following is a standard result on topological vector spaces. We include a proof for the sake of completeness. Lemma 3.4. Suppose V is a finite dimensional subspace of kΓ. Then V is closed with respect to the product topology on kΓ extending the discrete topology on k. Proof. In five easy steps. (i). Recall that in the discrete topology on k, a sequence converges if and only if it is eventually constant. Γ Γ (ii). Fix a ∈ k and write a = α(a, α) α. The product topology on k is such that a sequence {a ∈ kΓ} ⊆ V converges to a if and only if for each α ∈ Γ, there n exists Nα satisfying (an, α) = (a, α) for all n > Nα. 8 AARONLAUVEANDCHRISTOPHEREUTENAUER
(iii). Suppose that b1, . . . , bm is a basis for the finite dimensional subspace V ⊆ Γ k . Find elements α1,...,αm ∈ Γ satisfying
(b1, α1) (b2, α1) (bm, α1) (b1, α2) (b2, α2) (bm, α2) det . . . . = 0, . . .. . (b1, αm) (b2, αm) (bm, αm) which are guaranteed to exist by the independence of b1, . . . , bm. (n) (iv). Given {an}→ a as above, find constants si ∈ k satisfying (n) (n) (n) an = s1 b1 + s2 b2 + + sm bm (∀n). (3.5)
Put N = max {Nα1 ,...,Nαm }, with Nαi determined as in (ii). Let B be the matrix found in (iii). From (3.5), we have the system
(n) (b1, α1) (bm, α1) s1 (an, α1) . . . . . . .. . . = . . (b , α ) (b , α ) (n) (a , α ) 1 m m m sm n m Note that the right-hand side is constant for n > N. Specifically, it equals T [(a, α1) (a, αm)] . Since B is invertible, this system has a unique solution T [s1 sm] independent of n (for n > N). (v). We claim that
a = s1b1 + s2b2 + + smbm .
To see this, note that (3.5) reduces to an = s1b1 + + smbm for n > N, since the bi are linearly independent. It follows that an = a for all n > N and the lemma is proven.
3.3. Connes operators of rational elements. Given the preparatory results in Sections 3.1 and 3.2, we are ready to prove that Connes operators of rational ele- ments have finite rank.
Proof of Theorem 3.1 (Forward Implication). It is enough to induct on the (+, ×, ∗)-complexity of a rational series a ∈ kΓ. We check that if a ∈ kΓ then [F, a] has finite rank and that the finite rank condition is closed under +, × and ∗. (i): Closed under +. Suppose a = a′ + a′′ with [F, a′] and [F, a′′] finite rank. Then img [F, a] ⊆ img [F, a′]+ img [F, a′′], thus [F, a] has finite rank as well. (ii): Polynomials have finite rank. After (i) we may assume a is a monomial α ∈ Γ. Apply [F, a] to some e = ω eω ℘(ω),ω to get